DeepSeek

Use the deepseek provider for fast-agent's native DeepSeek Responses route. It uses a stateless Responses API over SSE and currently supports deepseek-v4-flash.

Setup

Create a key in the DeepSeek platform and set:

export DEEPSEEK_API_KEY="..."

Then run the built-in alias:

fast-agent go --model deepseek \
  --message "Explain why 37 × 41 equals 1517."

The explicit model string is:

deepseek.deepseek-v4-flash

The native provider rejects other model names. In particular, deepseek-chat, deepseek-reasoner, and deepseek-v4-pro are not native DeepSeek model strings in fast-agent.

Configuration

Only the API key is required:

deepseek:
  api_key: "${DEEPSEEK_API_KEY}"

The complete provider shape is:

deepseek:
  api_key: "${DEEPSEEK_API_KEY}"
  base_url: "https://api.deepseek.com"
  default_model: "deepseek-v4-flash"
  reasoning: "max"
  web_search:
    enabled: false
  # default_headers:
  #   X-Custom-Header: value

base_url, default_model, and default_headers are optional. A configured default_model must currently be deepseek-v4-flash.

Run fast-agent check after configuring credentials.

Reasoning

Reasoning defaults to max. Select an effort in the model string:

fast-agent go --model "deepseek?reasoning=none" --message "Answer directly."
fast-agent go --model "deepseek?reasoning=low" --message "Solve this problem."
fast-agent go --model "deepseek?reasoning=high" --message "Solve this problem."
fast-agent go --model "deepseek?reasoning=max" --message "Solve this problem."

Supported values are none, low, high, and max. DeepSeek returns reasoning separately from visible assistant text; fast-agent preserves it in the reasoning channel and replays it when continuing a tool-use turn.

max_output_tokens includes hidden reasoning. Leave enough output headroom when reasoning is enabled rather than treating the setting as a visible-text budget.

Tools, structured output, and web search

The native route supports:

function tools;
JSON Schema structured output;
provider-managed web search.

Enable web search in configuration:

deepseek:
  web_search:
    enabled: true

Or enable it for one model selection:

fast-agent go --model "deepseek?web_search=true"

DeepSeek currently accepts the web-search enablement toggle. Generic OpenAI-style search context, domain, and location options are not forwarded by the adapter.

When forcing a specific function with tool_choice, use reasoning=none. Automatic tool selection supports reasoning.

Stateless Responses behavior

DeepSeek's route differs from OpenAI's stateful Responses API:

requests use SSE; WebSocket transport is not supported;
server-side response storage and continuation are not used;
service tiers are not supported;
image, PDF, audio, video, and file inputs are not supported;
OpenAI-only request fields such as include, parallel_tool_calls, service_tier, and store are omitted.

Conversation and tool continuation still work because fast-agent sends the required history on each stateless request.

Hugging Face routes are separate

Aliases such as deepseek-hf, deepseek4-hf, and deepseek4pro-hf use Hugging Face Inference Providers, not the native DeepSeek API. They use HF_TOKEN, provider-specific routing, and their own capability metadata.

Use deepseek when you want the native DeepSeek Responses route. Use an hf. model string or a deepseek-*-hf alias when you want a Hugging Face-hosted route.

Model aliases

Model Alias	Maps to
`deepseek`	`deepseek.deepseek-v4-flash`
`deepseek-v4-flash`	`deepseek-v4-flash`

Official documentation

See Models Reference for the generated capability row, context limit, output limit, and supported input modalities.